You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
docs/usermanuals/en/reference/cli.md: broken example — jq -e '(.pii_severity.worst_tier // "none") != "high" and (.cross_split_leakage_pairs | length) == 0' audit.json gates training on high-tier P… (DOC-D3B04A-006) #221
The document says.jq -e '(.pii_severity.worst_tier // "none") != "high" and (.cross_split_leakage_pairs | length) == 0' audit.json gates training on high-tier PII or train/eval overlap.
What the code does.cross_split_leakage_pairs is the list of every split pair that was compared, including pairs with zero leaked rows. Reproduced on a clean train/validation corpus (0 leaked rows): the envelope had cross_split_leakage_pairs=['train__validation'], the documented jq printed false and exited 1. A replacement that reads leaked_rows_in_* from the on-disk report exited 0.
User impact. Under set -e the documented gate stops training on every clean multi-split corpus. Operators will disable or rewrite it.
Proposed correction. Replace the jq line with two checks: shell $ jq -e '(.pii_severity.worst_tier // "none") != "high"' audit.json $ jq -e '([.cross_split_overlap.pairs[]? | to_entries[] | select(.key | startswith("leaked_rows_in_")) | .value] | add // 0) == 0' audit/data_audit_report.json and explain that cross_split_leakage_pairs lists the compared split pairs, not only leaking ones.
Turkish mirror. the Turkish page has the same error; fix both in the same PR
Acceptance criteria
Each statement is re-checked against the code at fix time.
Documentation listed above is corrected in English and in its Turkish mirror in the same PR.
Every finding in this issue is resolved or explicitly closed with a reason.
Found by the October 2026 ForgeLM code and documentation review against commit cab2947. Code links point at that commit; line numbers may have moved since. Finding IDs (CR-…, DOC-…, TECH-…) are stable references for this backlog.
Activity